Model render profiles

The endpoint is a wire protocol, not a rendering profile. Claude, GPT, and Grok can all arrive on /v1/responses; pxpipe resolves geometry and vision billing from the model id.

A gateway may prefix that wire path with one routing segment and may qualify the model id with vendor segments. Both are recognized: /compat/chat/completions and /grok/v1/responses are read as the OpenAI shapes they are (one leading path segment is stripped), and the profile lookup tries the full id then its bare final segment, so moonshotai/kimi-k3 and workers-ai/@cf/moonshotai/kimi-k3 both resolve the same as kimi-k3 — vendor segments pick an upstream, not a geometry. Prefix recognition does not move routing — a provider-prefixed path still forwards to the passthrough upstream.

model rule	default	cell	columns	max height	evidence
claude-* / anthropic-*	yes	Spleen 5×8	312	728 px	established Claude suites
gpt-5.6-sol*	opt-in	JetBrains Mono 14px, 9×16	84	1954 px	7/8 exact, 0 inventions, gist and guard pass
grok-*	opt-in	JetBrains Mono 14px, 9×16	84	512 px	100/100 arith, 97/98 gist, 17/18 state; hex 0/15
other GPT/o-series	opt-in	Spleen 5×8	152	1932 px	conservative fallback

Every production path adds IDS rows to the image and an adjacent text factsheet for precision-critical strings. Recent and open protocol state remains native. Those guards reduce exact-string risk; they do not make image reading byte-safe.

Savings vs. cost

Native 14px is a legibility/quality profile, not a cost win on warm-cache traffic. Images do cache — a turn whose image prefix is already cached saves a little (dashboard: ~1,700 tokens on a measured Claude turn) — but images are large, so each time the prefix changes and the provider re-writes it, one cache_create costs several times the text it replaced. One measured turn sent 43,444 tokens versus 8,964 as text (−34,480), erasing ~20 warm turns of savings, i.e. net-negative. That is why the 14px profiles ship opt-in for legibility/quality; read cost off the dashboard's cache-aware number, and let the profitability gate skip transforms whose amortized create cost it predicts will lose.

Sol and Grok remain opt-in because their broader image-reading results do not match Claude. Enable them explicitly with PXPIPE_MODELS, for example:

PXPIPE_MODELS=claude-fable-5,gpt-5.6-sol

gpt-5.6-terra and other siblings do not inherit the Sol profile or allowlist.

Evidence: Sol results, Grok results.

Overrides

PXPIPE_GPT_PROFILES is a JSON map from model-id prefix to a partial profile. The longest prefix wins. Supported render fields are font, cellWBonus, cellHBonus, aa, grid, gridCols, colorCycle, markerScale, and markerRed; geometry fields are stripCols and maxHeightPx.

PXPIPE_GPT_PROFILES='{"gpt-5.6-sol":{"stripCols":120}}'

The profitability gate uses the same resolved profile as the renderer, so a style or geometry override cannot leave cost prediction on stale dimensions.

Per-content-class geometry

historyStripCols and historyStyle override stripCols and style for collapsed history only. The static slab and tool-result pages keep the dense geometry. Both are undefined in every shipped profile, so leaving them out changes nothing.

PXPIPE_GPT_PROFILES='{"claude-opus-5":{"historyStripCols":172,"historyStyle":{"font":"jetbrains-mono-14"}}}'

They exist because reading accuracy and rendering cost are not the same axis. Geometry is shared across a provider's models because the billing is, but verbatim recall is not. Measured on one 26-value battery — commit SHAs, 32-char md5 hashes, three-decimal ratios, currency to the cent, timestamped filenames — rendered at the shipped 312-column dense geometry:

reader	exact	silently wrong
Fable 5	25/26	1/26 (capitalisation)
Opus 5	3/26	10/26
Opus 5 at jetbrains-mono-14, 172 cols	100/100	0/100

So NOT-OCR.md's "unreliable at any size" holds at production density and not below it. The trade is real and the caller should make it: on 400 000 characters of live transcript, dense rendering is 5.28× cheaper than text and legible rendering 1.47×.

Set both fields together. A larger font at unchanged columns overshoots the provider's no-resize width, gets downscaled server-side, and loses exactly the legibility it was meant to buy.

Unmeasured families

An id that names a family pxpipe HAS measured, but does not match any of that family's profiles — an unmeasured Gemini or Grok sibling — is refused by isMisresolvedModelId, and the request passes through as text. Such an id would otherwise fall through to DEFAULT_GPT_PROFILE and be gated with OpenAI's tile math, i.e. priced with the wrong provider's formula.

An id that names no known family at all (anything reached through a gateway's OpenAI-compatible route — Kimi K3 on Cloudflare is the example this repo has run) is NOT refused. It resolves to DEFAULT_GPT_PROFILE and is gated with OpenAI's tile math as an approximation. PXPIPE_MODELS is the only gate: listing such a model is the operator asserting it is worth imaging. Two consequences worth stating plainly:

• The savings figure is an approximation, not a measurement. If tile math overstates the model's real image cost, the gate declines transforms that would have paid off; if it understates, you take a small loss.
• Nothing checks that the model can accept images at all. A text-only model will reject an imaged request outright — a 400 on the first compressed call, not a silent degradation.

To replace the approximation with real numbers, declare the geometry and vision cost yourself:

PXPIPE_MODELS=claude-fable-5,kimi-k3
PXPIPE_GPT_PROFILES='{"kimi-k3":{"vision":{"regime":"mpix","tokensPerMegapixel":1000},"stripCols":152,"maxHeightPx":1932}}'

The tokensPerMegapixel above is a placeholder, not a measurement — read the real image-token rule off your provider's billing docs, or measure it by sending one known-size image and reading the reported input tokens. A wrong number here does not corrupt output; it corrupts the profitability gate and the dashboard's savings figure, which is worse, because both will look confident.

Declaring a cost says nothing about whether the model can read dense imaged text. Verify that separately before trusting a compressed session — see eval/ for the harnesses used on the shipped profiles.